It seems that avoiding to create local aliases for the self._rlock and
self._wlock methods prevents the strange errors on i686 related to
semaphore locking and unlocking.

One example of failing builds is firefox-esr running gyp in an i686 environment.

--- Lib/multiprocessing/queues.py	2017-09-16 19:38:35.000000000 +0200
+++ Lib/multiprocessing/queues.py	2017-10-24 19:49:06.291351206 +0200
@@ -369,13 +369,11 @@
 
     def _make_methods(self):
-        recv = self._reader.recv
-        racquire, rrelease = self._rlock.acquire, self._rlock.release
         def get():
-            racquire()
-            try:
-                return recv()
-            finally:
-                rrelease()
+            self._rlock.acquire()
+            try:
+                return self._reader.recv()
+            finally:
+                self._rlock.release()
         self.get = get
 
         if self._wlock is None:
@@ -383,11 +382,9 @@
             self.put = self._writer.send
         else:
-            send = self._writer.send
-            wacquire, wrelease = self._wlock.acquire, self._wlock.release
             def put(obj):
-                wacquire()
-                try:
-                    return send(obj)
-                finally:
-                    wrelease()
+                self._wlock.acquire()
+                try:
+                    return self._writer.send(obj)
+                finally:
+                    self._wlock.release()
             self.put = put
--- Modules/_multiprocessing/semaphore.c	2017-09-16 19:38:35.000000000 +0200
+++ Modules/_multiprocessing/semaphore.c	2017-10-28 10:49:56.944993401 +0200
@@ -378,7 +378,7 @@
             }
         }
 #else
-        int sval;
+        int sval = -1;
 
         /* This check is not an absolute guarantee that the semaphore
            does not rise above maxvalue. */